Machine learning in molecular ecology

Unsupervised machine learning for species delimitation, integrative taxonomy, and biodiversity conservation

Integrative taxonomy, combining data from multiple axes of biologically relevant variation, is a major goal of systematics. Ideally, such taxonomies will derive from similarly integrative species-delimitation analyses. Yet, most current methods rely solely or primarily on molecular data, with other layers often incorporated only in a post hoc qualitative or comparative manner. A major limitation is the difficulty of devising quantitative parametric models linking different datasets in a unified ecological and evolutionary framework. Machine Learning (ML) methods offer flexibility in this arena by easily learning high-dimensional associations between observations (e.g., individual specimens) across a wide array of input features (e.g., genetics, geography, environment, and phenotype) to delimit statistically meaningful clusters. Here, I implement an unsupervised method using Self-Organizing (or "Kohonen") Maps (SOMs) for such purposes. Recent extensions called "SuperSOMs" can integrate multiple layers, each of which exerts independent influence on a two-dimensional output grid via empirically estimated weights. The grid cells are then delimited into K distinct units that can be interpreted as species or other entities. I show empirical examples in salamanders (Desmognathus) and snakes (Storeria) with layers representing alleles, space, climate, and traits. Simulations reveal that the SuperSOM approach can detect K = 1, tends not to over-split, reflects contributions from all layers, and limits large layers (e.g., genetic matrices) from overwhelming other datasets, desirable properties addressing major concerns from previous studies. Finally, I suggest that these and similar methods could integrate conservation-relevant layers such as population trends and human encroachment to delimit management units from an explicitly quantitative framework grounded in the ecology and evolution of species limits and boundaries.

Deep Learning in Population Genetics

Population genetics is transitioning into a data-driven discipline thanks to the availability of large-scale genomic data and the need to study increasingly complex evolutionary scenarios. With likelihood and Bayesian approaches becoming either intractable or computationally unfeasible, machine learning, and in particular deep learning, algorithms are emerging as popular techniques for population genetic inferences. These approaches rely on algorithms that learn non-linear relationships between the input data and the model parameters being estimated through representation learning from training data sets. Deep learning algorithms currently employed in the field comprise discriminative and generative models with fully connected, convolutional, or recurrent layers. Additionally, a wide range of powerful simulators to generate training data under complex scenarios are now available. The application of deep learning to empirical data sets mostly replicates previous findings of demography reconstruction and signals of natural selection in model organisms. To showcase the feasibility of deep learning to tackle new challenges, we designed a branched architecture to detect signals of recent balancing selection from temporal haplotypic data, which exhibited good predictive performance on simulated data. Investigations on the interpretability of neural networks, their robustness to uncertain training data, and creative representation of population genetic data, will provide further opportunities for technological advancements in the field.

Habitat connectivity and host relatedness influence virus spread across an urbanising landscape in a fragmentation-sensitive carnivore

Spatially heterogeneous landscape factors such as urbanisation can have substantial effects on the severity and spread of wildlife diseases. However, research linking patterns of pathogen transmission to landscape features remains rare. Using a combination of phylogeographic and machine learning approaches, we tested the influence of landscape and host factors on feline immunodeficiency virus (FIV_Lru) genetic variation and spread among bobcats (Lynx rufus) sampled from coastal southern California. We found evidence for increased rates of FIV_Lru lineage spread through areas of higher vegetation density. Furthermore, single-nucleotide polymorphism (SNP) variation among FIV_Lru sequences was associated with host genetic distances and geographic location, with FIV_Lru genetic discontinuities precisely correlating with known urban barriers to host dispersal. An effect of forest land cover on FIV_Lru SNP variation was likely attributable to host population structure and differences in forest land cover between different populations. Taken together, these results suggest that the spread of FIV_Lru is constrained by large-scale urban barriers to host movement. Although urbanisation at fine spatial scales did not appear to directly influence virus transmission or spread, we found evidence that viruses transmit and spread more quickly through areas containing higher proportions of natural habitat. These multiple lines of evidence demonstrate how urbanisation can change patterns of contact-dependent pathogen transmission and provide insights into how continued urban development may influence the incidence and management of wildlife disease.

eDNAssay: a machine learning tool that accurately predicts qPCR cross-amplification

Environmental DNA (eDNA) sampling is a highly sensitive and cost-effective technique for wildlife monitoring, notably through the use of qPCR assays. However, it can be difficult to ensure assay specificity when many closely related species cooccur. In theory, specificity may be assessed in silico by determining whether assay oligonucleotides have enough base-pair mismatches with nontarget sequences to preclude amplification. However, the mismatch qualities required are poorly understood, making in silico assessments difficult and often necessitating extensive in vitro testing-typically the greatest bottleneck in assay development. Increasing the accuracy of in silico assessments would therefore streamline the assay development process. In this study, we paired 10 qPCR assays with 82 synthetic gene fragments for 530 specificity tests using SYBR Green intercalating dye (n = 262) and TaqMan hydrolysis probes (n = 268). Test results were used to train random forest classifiers to predict amplification. The primer-only model (SYBR Green-based) and full-assay model (TaqMan probe-based) were 99.6% and 100% accurate, respectively, in cross-validation. We further assessed model performance using six independent assays not used in model training. In these tests the primer-only model was 92.4% accurate (n = 119) and the full-assay model was 96.5% accurate (n = 144). The high performance achieved by these models makes it possible for eDNA practitioners to more quickly and confidently develop assays specific to the intended target. Practitioners can access the full-assay model via eDNAssay (https://NationalGenomicsCenter.shinyapps.io/eDNAssay), a user-friendly online tool for predicting qPCR cross-amplification.

Machine Learning models identify gene predictors of waggle dance behaviour in honeybees

The molecular characterisation of complex behaviours is a challenging task as a range of different factors are often involved to produce the observed phenotype. An established approach is to look at the overall levels of expression of brain genes - or 'neurogenomics' - to select the best candidates that associate with patterns of interest. However, traditional neurogenomic analyses have some well-known limitations; above all, the usually limited number of biological replicates compared to the number of genes tested - known as "curse of dimensionality". In this study we implemented a Machine Learning (ML) approach that can be used as a complement to more established methods of transcriptomic analyses. We tested three supervised learning algorithms (Random Forests, Lasso and Elastic net Regularized Generalized Linear Model, and Support Vector Machine) for their performance in the characterization of transcriptomic patterns and identification of genes associated with honeybee waggle dance. We then intersected the results of these analyses with traditional outputs of differential gene expression analyses and identified two promising candidates for the neural regulation of the waggle dance: boss and hnRNP A1. Overall, our study demonstrates the application of Machine Learning to analyse transcriptomics data and identify candidate genes underlying social behaviour. This approach has great potential for application to a wide range of different scenarios in evolutionary ecology, when investigating the genomic basis for complex phenotypic traits and can present some clear advantages compared to the established tools of gene expression analysis, making it a valuable complement for future studies.